The current generation of AI products is astonishing on the inside and, too often, underdesigned on the outside. The models can reason, write, and code at a level that felt impossible two years ago — yet the experience wrapped around them frequently ignores decades of hard-won UX craft. Capability has raced ahead of usability.
I spend my days auditing product experiences, so let me be specific rather than vague. I'll use Claude — one of the most capable assistants available — as a running case study. Not because it's the worst (it isn't), but because the gap between how good the model is and how under-designed the experience around it can feel is exactly the point. Almost everything here applies across the category.
1. The blank-canvas problem
Most AI products, Claude included, still greet you with an empty text box and a blinking cursor. For a power user, that's freedom. For everyone else, it's the tyranny of the blank page: what do I even type? A model that can do thousands of things surfaces none of them. Good UX lowers the cost of getting started; a blank box maximizes it.
The fix
Intent-driven entry points — example prompts tied to real jobs, templates, and contextual suggestions — so users discover capability instead of guessing at it.
2. Discoverability of features
AI tools ship capabilities at a breathless pace — projects, memory, styles, tools, connectors — but they're often buried behind icons and menus a typical user never opens. In Claude and its peers, some of the most powerful features are effectively invisible unless you already know they exist. A feature nobody can find is, functionally, a feature that doesn't exist.
The fix
Progressive disclosure and in-context nudges that reveal the right capability at the moment it's useful — not a settings panel the user has to excavate.
3. Trust, transparency, and the black box
AI output is probabilistic and occasionally wrong, yet the interface often presents it with the flat confidence of a calculator result. Where did this come from? How sure is it? What did it actually use — my files, the web, its memory? When products don't answer those questions visibly, trust erodes — especially for the enterprise users who matter most.
The fix
Design for explainability: show sources, states, and boundaries. Make it obvious what the system knows, what it's doing, and where it might be wrong.
4. Loading, empty, and error states
Traditional software sweats over these states. AI products, moving fast, often skip them — a long silent spinner, a truncated answer, or a hard failure with no graceful path back. The absence of thoughtful loading, empty, and error states is one of the clearest tells that UX was an afterthought.
The fix
Streaming feedback that shows real progress, inline controls to edit or regenerate, and honest, recoverable error states.
Why I'm not just picking on one product
None of this is unique to Claude — it's the norm across the category, from the biggest labs down. That's the real headline: the industry keeps treating UX as a wrapper instead of the product. Building a powerful model is an engineering achievement; building something people can actually use well is a design one. As models converge in raw capability, the experience is where products will win or lose.
To be fair, these teams ship at a pace most companies can't match, and the underlying tools are genuinely remarkable — which is exactly why the UX gap is worth closing. A modest amount of design rigor would compound enormously.
If you're building an AI product and want the experience to match the intelligence under the hood, let's talk.
